# 017 role-heatmap — RESULT (harvested tick 18, 2026-08-02, tier T3-exploratory)

**Verdict: the order code is STRICTLY FILLER-LOCAL at every depth — it never broadcasts.
The "broadcast before pooling" hypothesis (from 014's L24 cliff) is REFUTED; and role is
chance in all 72 cells.**

## Heatmap (pooled AUC, cross-construction + lexical holdout, block-rule eligibility)
- **Surface (order) target**: fillerA/fillerB spans 0.86-0.96 from L0 through L24n
  (0.94/0.93 at the post-LN pooling input); EVERY other position class — lang, first,
  predicate, other-content, final, eos — at 0.47-0.53 at ALL layers; **allpool ~0.50-0.52**.
- **Role target**: 0.48-0.55 across all 9 classes × 8 layers (max 0.55, pred@L16, CI at
  chance). Nothing.

## Interpretation (scoped)
1. Order information rides exclusively on the moving tokens themselves at every stage —
   summary positions (lang/eos/final) never accumulate it, even by the pooling input.
2. This REFRAMES 014's L24 patch failure: role/order isn't "broadcast elsewhere by L24"
   (nothing spreads); rather, a last-layer local patch has no remaining propagation
   layers to amplify the swap, and the pooled mean dilutes 2-4 swapped tokens below the
   decoder's decision basin. Locality holds; commitment is about amplification-time.
3. New tension worth a writeup note: the pooled vector's ORDER content is ~chance under
   this strict transfer cell, yet the decoder reconstructs order near-verbatim — the
   decoder's order-reading of z must be nonlinear/lexically-conditioned (word-identity ×
   position products surviving the mean), consistent with 016's "nonlinear families see
   the surface code" and the C4 decode-vs-probe theme.
4. 018's companion result (function words = conduits, not stores) + this map = the arc's
   position-level picture is complete: everything role/order-like is carried by the
   argument tokens, locally, in construction-specific format.

## Brier (frozen priors): (a) 0.85 → T (0.0225); (b) 0.60 → F (0.36 — the registered
hypothesis refuted); (c) 0.10 → F (0.01); (d) ~0.35 → F (0.1225). Mean **0.129**.

## Incidents: first run FAILED on prop-imbalance (block-rule recurrence); tick-patched
and relaunched same prereg (machinery fix, logged t17). Wall 330s post-fix.